def test():
import numpy as np
video_shape = (16, 112, 112)
data = DataLayer("data/tinytraindb.lmdb", video_shape, 16, verbose=True)
data2 = DataLayer("data/tinyvaldb.lmdb", video_shape, 16, verbose=True)
synch_data = []
# Retrieve data synchronously as a reference for correctness
fetcher = DataFetcher("data/tinytraindb.lmdb",
video_shape,
16,
dtype='float32')
for i in range(10):
X, y, epoch = fetcher.load_data()
synch_data.append(X)
#print X
for i in range(10):
data.load_batch()
data2.load_batch()
tic = time.time()
# Do some work
a = np.random.randn(400, 600).dot(np.random.randn(600, 400))
#print np.linalg.norm(a)
toc = time.time()
print "Work took %0.6f seconds" % (toc - tic)
assert np.linalg.norm(data.X.get_value(borrow=True) -
synch_data[i]) < 1e-8
def __init__(self,
db_name,
video_shape,
mem_batch_size,
verbose=False,
buffer_size=6):
self.fetcher = DataFetcher(db_name,
video_shape,
mem_batch_size,
dtype=theano.config.floatX)
self.batch_size = mem_batch_size
self.video_shape = video_shape
self.current_batch = 0
self.verbose = verbose
# Could manage with a buffer size of exactly 2, but need to change
# the interprocess communication somewhat
assert buffer_size > 2
X = np.empty((mem_batch_size, 3) + video_shape,
dtype=theano.config.floatX)
y = np.empty((mem_batch_size, ), dtype=theano.config.floatX)
self.shared_data = theano.shared(X, borrow=True)
self.shared_label = theano.shared(y, borrow=True)
self.X = self.shared_data
self.y = T.cast(self.shared_label, 'int32')
# Create shared memory object for async loading
X_shared_array_base = multiprocessing.Array(
ctypes.c_float,
buffer_size * self.batch_size * 3 * np.prod(video_shape))
X_shared_array = np.ctypeslib.as_array(X_shared_array_base.get_obj())
self.X_shared_array = X_shared_array.reshape(buffer_size,
self.batch_size, 3,
*video_shape)
y_shared_array_base = multiprocessing.Array(
ctypes.c_float, buffer_size * self.batch_size)
y_shared_array = np.ctypeslib.as_array(y_shared_array_base.get_obj())
self.y_shared_array = y_shared_array.reshape(buffer_size,
self.batch_size)
# Start up worker process
self.queue = multiprocessing.Queue(maxsize=buffer_size - 2)
self.worker = multiprocessing.Process(
target=fetcher_loop,
args=(self.fetcher, self.X_shared_array, self.y_shared_array,
self.queue))
self.worker.start()
def test():
import numpy as np
video_shape = (16,112,112)
data = DataLayer("data/tinytraindb.lmdb",video_shape,16,verbose=True)
data2 = DataLayer("data/tinyvaldb.lmdb",video_shape,16,verbose=True)
synch_data = []
# Retrieve data synchronously as a reference for correctness
fetcher = DataFetcher("data/tinytraindb.lmdb",video_shape,16,dtype='float32')
for i in range(10):
X, y, epoch = fetcher.load_data()
synch_data.append(X)
#print X
for i in range(10):
data.load_batch()
data2.load_batch()
tic = time.time()
# Do some work
a = np.random.randn(400,600).dot(np.random.randn(600,400))
#print np.linalg.norm(a)
toc = time.time()
print "Work took %0.6f seconds" % (toc - tic)
assert np.linalg.norm(data.X.get_value(borrow=True) - synch_data[i]) < 1e-8
def evaluate_3d_conv():
theano.config.exception_verbosity = "high"
theano.config.optimizer = 'None'
rng = np.random.RandomState(234)
TT, HH, WW = 16,240,320
N = 10
num_classes = 5
batch_size = 1
num_filters = 4
num_channels = 3
if len(sys.argv) > 1:
fetcher = DataFetcher("data/tinyvideodb.lmdb")
X, y = fetcher.load_data(10,(16,240,320))
y /= 21
else:
X = np.random.randint(-127,127,size=(N,3,16,240,320)).astype(theano.config.floatX)
y = np.random.randint(0,num_classes,size=(N,))
X_train = theano.shared(X.astype('float32'), borrow=True)
y_train = theano.shared(y.astype('int32'), borrow=True)
print y_train.get_value()
params = []
x = dtensor5('x')
y = T.ivector('y')
FT, FH, FW = 5, 5, 5
###########################################################################
# CONV-RELU-POOL (Layer 1)
###########################################################################
conv1 = ConvLayer(x,num_channels,num_filters,(FT,FH,FW),(TT,HH,WW),batch_size,relu,
layer_name="Conv1")
params += conv1.params
pool1 = PoolLayer(conv1.output,(2,2,2))
###########################################################################
# CONV-RELU-POOL (Layer 2)
###########################################################################
conv2 = ConvLayer(pool1.output,num_filters,num_filters,
(FT,FH,FW),
(TT/2,HH/2,WW/2),
batch_size,
relu,
layer_name="Conv2")
params += conv2.params
pool2 = PoolLayer(conv2.output,(2,2,2))
###########################################################################
# FULLY-CONNECTED (Layer 3)
###########################################################################
out_dim = num_filters*TT*HH*WW/64
num_hidden = 64
fc3 = HiddenLayer(pool2.output.flatten(ndim=2),out_dim,num_hidden,relu)
params += fc3.params
###########################################################################
# SOFTMAX (Layer 4)
###########################################################################
softmax = LogRegr(fc3.output,num_hidden,num_classes,relu,rng)
params += softmax.params
reg = 0.01
cost = softmax.negative_log_likelihood(y) + reg*T.sum(softmax.W*softmax.W)
# compute the gradient of cost with respect to theta (sotred in params)
# the resulting gradients will be stored in a list gparams
gparams = [T.grad(cost, param) for param in params]
learning_rate = 1e-5
updates = [
(param, param - learning_rate * gparam)
for param, gparam in zip(params, gparams)
]
index = T.lscalar()
train_model = theano.function(
inputs=[index],
outputs=cost,
updates=updates,
givens={
x: X_train[index * batch_size: (index + 1) * batch_size],
y: y_train[index * batch_size: (index + 1) * batch_size]
}
)
for k in range(10):
tic = time.time()
cost = train_model(k % (N/batch_size))
toc = time.time()
print cost, "(%0.4f seconds)" % (toc - tic)
